Impact of Imaging Modality on AI-Based Detection of Incidental Maxillary Sinus Pathology: Comparison of Panoramic Radiography and CBCT.

Background/

Objectives

The objective of our study was to compare the diagnostic performance of a popular, commercial dental artificial intelligence (AI) platform (Diagnocat, DGNCT LLC, Miami, FL, USA) for detecting maxillary sinus abnormalities on paired panoramic radiographs (OPG) and cone-beam computed tomography (CBCT) acquired in the same patients, and to examine whether lesion conspicuity predicts correct AI decisions.

Methods

In this retrospective paired study, 166 patients contributed 332 maxillary sinuses with OPG and CBCT performed ≤30 days apart. The reference standard was consensus CBCT reading by two observers with third-reader arbitration. The index test was the AI's sinus-level binary output (any abnormality). Accuracy, precision, recall, and F1 score were estimated with patient-clustered 95% bootstrap CIs; secondary analyses assessed category-specific performance and the effect of mucosal thickness and polyp/cyst volume.

Results

Our evaluation showed that the platform's performance depended on modality. On CBCT, the accuracy was 69.88% (64.76-74.70%), precision was 87.83% (81.58-93.33%), recall was 54.01% (46.74-61.17%), and F1 score was 66.89% (60.34-72.84%). On OPG, the accuracy was 50.6% (44.58-55.41%), precision was 67.80% (55.38-79.66%), recall was 21.39% (15.62-27.32%), and F1 score 32.52% (24.79-39.69%). On CBCT, higher mucosal thickness and larger polyp/cyst volume strongly predicted correct AI calls; no such effect was seen with OPG.

Conclusions

In conclusion, the evaluated AI showed high precision but only moderate recall on CBCT and unreliable performance on OPG. Outputs must be interpreted by a professional; AI alerts on OPG should not guide management without CBCT confirmation.